Import

import numpy as np 
import pandas as pd 
import matplotlib
import matplotlib.pyplot as plt
import seaborn as sns
import os
from google.colab import drive
drive.mount('/content/drive')
from sklearn.model_selection import train_test_split

Mounted at /content/drive

!pip install --quiet shap==0.39.0
!pip install --quiet fastai==2.2.5

     |████████████████████████████████| 358kB 9.1MB/s 
  Building wheel for shap (setup.py) ... done
     |████████████████████████████████| 194kB 7.8MB/s 
     |████████████████████████████████| 12.8MB 246kB/s 
     |████████████████████████████████| 776.8MB 24kB/s 
     |████████████████████████████████| 61kB 9.4MB/s 
ERROR: torchtext 0.9.0 has requirement torch==1.8.0, but you'll have torch 1.7.1 which is incompatible.

import sys 
sys.path.append('/content/drive/MyDrive/Tesi_magistrale/Tesi_magistrale')
import mf

Dataset import

data = pd.read_csv('/content/drive/MyDrive/Tesi_magistrale/Dataset/IEEE/Output/data.csv')
data = mf.new_processing(data)
data = pd.get_dummies(data)

Model import

XGBoost

pd.options.mode.chained_assignment = None
X_test = mf.get_set('test_id', data, 'xgboost_new')

Importante: se il modello è trainato su GPU, quando si carica bisogna attivare il runtime con GPU, sennò il modello non vine caricato nel modo corretto.

load = True
if load:
  from joblib import load
  from xgboost import XGBClassifier

  xgb = load("/content/drive/MyDrive/Tesi_magistrale/Dataset/IEEE/Output/xgboost_new/xgb_np.joblib.dat")

SHAP

Idea

A prediction can be explained by assuming that each feature value of the instance is a “player” in a game where the prediction is the payout. Shapley values – a method from coalitional game theory – tells us how to fairly distribute the “payout” among the features.

Definition

The Shapley value is the average marginal contribution of a feature value across all possible coalitions.

The Shapley value is NOT the difference in prediction when we would remove the feature from the model.

Interpretation

Given the current set of feature values, the contribution of a feature value to the difference between the actual prediction and the mean prediction is the estimated Shapley value.

Theoretical algorithm

Per determinare il valore di Shapley della feature x_j, si elimina la feature dal dataset, e delle altre si fanno tutte le possibili estrazioni, andando a fare la predizione solo con le feature selezionate e aggiungendo x_j. Facendo la differenza tra ogni predizione con e senza x_j si ottiene il contributo marginale di x_j. Per fare la predizione con il modello quando delle feature sono state eliminate, si rimpiazzano i loro valori con valori random (ovviamente tra quelli delle feature rimosse) estratti dal dataset.

Implementation

SHAP turns the Shapley values method into an optimization problem and uses a special kernel function to measure proximity of data instances. The results of SHAP are sparse

Riferimenti

Repo github: https://github.com/slundberg/shap

Tutorial (vedere parte per RF): https://towardsdatascience.com/explain-any-models-with-the-shap-values-use-the-kernelexplainer-79de9464897a

Grafici con codice funzionante: https://shap.readthedocs.io/en/latest/example_notebooks/tabular_examples/tree_based_models/Census%20income%20classification%20with%20LightGBM.html?highlight=classification

def change_feature_names(xgb, X_test):
  booster = xgb.get_booster()
  features = []
  for i in range(X_test.shape[1]):
    string = 'f'+str(i)
    features.append(string)
  booster.feature_names = features
  return 

change_feature_names(xgb, X_test)

import shap
shap.initjs()
explainer = shap.KernelExplainer(xgb.predict_proba, shap.sample(X_test,100))

def get_fraud_id(model, X_test, soglia = 0.3, punti = 20):
  import random
  predictions = model.predict_proba(X_test.to_numpy())[:,1]
  lista = [i for i,x in enumerate(predictions) if x>soglia]
  return random.sample(lista, punti)

indici = get_fraud_id(xgb, X_test)
lista = np.random.randint(100, 10000, size = 100).tolist()
# lista = np.arange(100,120).tolist()

z = X_test.iloc[indici+lista]
# kernel_shap = explainer.shap_values(z) #sono i valori di shapley

# np.save('shap_values_all',kernel_shap)

kernel_shap = np.load('/content/drive/MyDrive/Tesi_magistrale/Dataset/IEEE/Output/xgboost_new/shap/shap_values_all.npy')

predictions = xgb.predict_proba(z.to_numpy())

feature_names = X_test.columns

shap.initjs()
shap.force_plot(explainer.expected_value[1], kernel_shap[1][0,:], z.to_numpy()[0], feature_names=feature_names)

Visualization omitted, Javascript library not loaded!
Have you run `initjs()` in this notebook? If this notebook was from another user you must also trust this notebook (File -> Trust notebook). If you are viewing this notebook on github the Javascript has been stripped for security. If you are using JupyterLab this error is because a JupyterLab extension has not yet been written.

shap.initjs()
shap.force_plot(explainer.expected_value[1], kernel_shap[1][1,:], z.to_numpy()[1], feature_names=feature_names)

Visualization omitted, Javascript library not loaded!
Have you run `initjs()` in this notebook? If this notebook was from another user you must also trust this notebook (File -> Trust notebook). If you are viewing this notebook on github the Javascript has been stripped for security. If you are using JupyterLab this error is because a JupyterLab extension has not yet been written.

shap.initjs()
shap.force_plot(explainer.expected_value[1], kernel_shap[1][2,:], z.to_numpy()[2], feature_names=feature_names)

Visualization omitted, Javascript library not loaded!
Have you run `initjs()` in this notebook? If this notebook was from another user you must also trust this notebook (File -> Trust notebook). If you are viewing this notebook on github the Javascript has been stripped for security. If you are using JupyterLab this error is because a JupyterLab extension has not yet been written.

shap.initjs()
shap.force_plot(explainer.expected_value[1], kernel_shap[1][3,:], z.to_numpy()[3], feature_names=feature_names)

Visualization omitted, Javascript library not loaded!
Have you run `initjs()` in this notebook? If this notebook was from another user you must also trust this notebook (File -> Trust notebook). If you are viewing this notebook on github the Javascript has been stripped for security. If you are using JupyterLab this error is because a JupyterLab extension has not yet been written.

shap.initjs()
shap.force_plot(explainer.expected_value[1], kernel_shap[1][8,:], z.to_numpy()[8], feature_names=feature_names)

Visualization omitted, Javascript library not loaded!
Have you run `initjs()` in this notebook? If this notebook was from another user you must also trust this notebook (File -> Trust notebook). If you are viewing this notebook on github the Javascript has been stripped for security. If you are using JupyterLab this error is because a JupyterLab extension has not yet been written.

shap.initjs()
shap.force_plot(explainer.expected_value[1], kernel_shap[1][12,:], z.to_numpy()[12], feature_names=feature_names)

Visualization omitted, Javascript library not loaded!
Have you run `initjs()` in this notebook? If this notebook was from another user you must also trust this notebook (File -> Trust notebook). If you are viewing this notebook on github the Javascript has been stripped for security. If you are using JupyterLab this error is because a JupyterLab extension has not yet been written.

shap.initjs()
shap.force_plot(explainer.expected_value[1], kernel_shap[1][13,:], z.to_numpy()[13], feature_names=feature_names)

Visualization omitted, Javascript library not loaded!
Have you run `initjs()` in this notebook? If this notebook was from another user you must also trust this notebook (File -> Trust notebook). If you are viewing this notebook on github the Javascript has been stripped for security. If you are using JupyterLab this error is because a JupyterLab extension has not yet been written.

shap.initjs()
shap.force_plot(explainer.expected_value[1], kernel_shap[1][14,:], z.to_numpy()[14], feature_names=feature_names)

Visualization omitted, Javascript library not loaded!
Have you run `initjs()` in this notebook? If this notebook was from another user you must also trust this notebook (File -> Trust notebook). If you are viewing this notebook on github the Javascript has been stripped for security. If you are using JupyterLab this error is because a JupyterLab extension has not yet been written.

shap.initjs()
shap.force_plot(explainer.expected_value[1], kernel_shap[1][4,:], z.to_numpy()[4], feature_names=feature_names)

Visualization omitted, Javascript library not loaded!
Have you run `initjs()` in this notebook? If this notebook was from another user you must also trust this notebook (File -> Trust notebook). If you are viewing this notebook on github the Javascript has been stripped for security. If you are using JupyterLab this error is because a JupyterLab extension has not yet been written.

shap.initjs()
shap.force_plot(explainer.expected_value[1], kernel_shap[1][5,:], z.to_numpy()[5], feature_names=feature_names)

Visualization omitted, Javascript library not loaded!
Have you run `initjs()` in this notebook? If this notebook was from another user you must also trust this notebook (File -> Trust notebook). If you are viewing this notebook on github the Javascript has been stripped for security. If you are using JupyterLab this error is because a JupyterLab extension has not yet been written.

shap.initjs()
shap.force_plot(explainer.expected_value[1], kernel_shap[1][-1,:], z.to_numpy()[-1], feature_names=feature_names)

Visualization omitted, Javascript library not loaded!
Have you run `initjs()` in this notebook? If this notebook was from another user you must also trust this notebook (File -> Trust notebook). If you are viewing this notebook on github the Javascript has been stripped for security. If you are using JupyterLab this error is because a JupyterLab extension has not yet been written.

shap.initjs()
shap.force_plot(explainer.expected_value[1], kernel_shap[1][-2,:], z.to_numpy()[-2], feature_names=feature_names)

Visualization omitted, Javascript library not loaded!
Have you run `initjs()` in this notebook? If this notebook was from another user you must also trust this notebook (File -> Trust notebook). If you are viewing this notebook on github the Javascript has been stripped for security. If you are using JupyterLab this error is because a JupyterLab extension has not yet been written.

shap.initjs()
shap.force_plot(explainer.expected_value[1], kernel_shap[1][-3,:], z.to_numpy()[-3], feature_names=feature_names)

Visualization omitted, Javascript library not loaded!
Have you run `initjs()` in this notebook? If this notebook was from another user you must also trust this notebook (File -> Trust notebook). If you are viewing this notebook on github the Javascript has been stripped for security. If you are using JupyterLab this error is because a JupyterLab extension has not yet been written.

shap.initjs()
shap.force_plot(explainer.expected_value[1], kernel_shap[1][-6,:], z.to_numpy()[-6], feature_names=feature_names)

Visualization omitted, Javascript library not loaded!
Have you run `initjs()` in this notebook? If this notebook was from another user you must also trust this notebook (File -> Trust notebook). If you are viewing this notebook on github the Javascript has been stripped for security. If you are using JupyterLab this error is because a JupyterLab extension has not yet been written.

shap.initjs()
shap.force_plot(explainer.expected_value[1], kernel_shap[1][-11,:], z.to_numpy()[-11], feature_names=feature_names)

Visualization omitted, Javascript library not loaded!
Have you run `initjs()` in this notebook? If this notebook was from another user you must also trust this notebook (File -> Trust notebook). If you are viewing this notebook on github the Javascript has been stripped for security. If you are using JupyterLab this error is because a JupyterLab extension has not yet been written.

shap.summary_plot(kernel_shap[1], z.to_numpy(), feature_names=feature_names)

Decision plot: https://shap.readthedocs.io/en/latest/example_notebooks/api_examples/plots/decision_plot.html?highlight=plot

shap.decision_plot(explainer.expected_value[1], kernel_shap[1][0,:], z.to_numpy()[0], feature_names=feature_names.to_list())

shap.decision_plot(explainer.expected_value[1], kernel_shap[1][1,:], z.to_numpy()[1], feature_names=feature_names.to_list())

shap.decision_plot(explainer.expected_value[1], kernel_shap[1][2,:], z.to_numpy()[2], feature_names=feature_names.to_list())

shap.decision_plot(explainer.expected_value[1], kernel_shap[1][3,:],z.to_numpy()[3], feature_names=feature_names.to_list())

shap.decision_plot(explainer.expected_value[1], kernel_shap[1][-1,:], z.to_numpy()[-1], feature_names=feature_names.to_list())